[BUG] Lost provider mount replies leave allocations without cleanup custody #13

Closed
opened 2026-09-30 05:10:49 +02:00 by erikinkinen · 0 comments
Owner

Summary

A lost or malformed filesystem-provider mount reply can leave an allocated provider mount without a Vfsd-owned cleanup identity. This is a source-derived reliability and resource-lifetime defect; no executed compiler, fault-injection or VM reproduction is claimed.

Reproduction

Starting from Vfsd revision e08309b155, have a provider commit its mount allocation and suppress or corrupt the corresponding private mount response. Vfsd's mount_provider returns an error before mount_fs inserts the reserved namespace record. The provider mount ID exists only in the lost response. The old mount family supplies neither a caller-known creation key nor an original-key retirement operation. Repeat the fault to consume provider mount capacity. This is a proposed transport fault-injection sequence derived from source, not a claim that it was run.

Expected and actual behavior

Expected: mount creation is counted before submission and retains a caller-known identity that can retire the original provider allocation even if no reply is received. Parent traversal custody must also remain retained until cleanup is positively complete.

Actual source behavior: the vacant namespace reservation is abandoned on the fallible mount call; no record retains a possible provider allocation and no private cleanup family can address it without the returned mount ID.

Environment and identity

  • Component: Vfsd e08309b155, original private provider mount family 0x620.
  • Integration source graph: cbca80fc5fb7eb00c783d7e4ff3af1cd9f471890.
  • Affected provider allocation paths: Ramfsd create_instance and ext/FAT state mount installation after media validation.
  • Native authority: the existing explicitly delegated filesystem-provider endpoint; no arbitrary client endpoint or supplied numeric ID grants that authority.
  • No image, VM or hardware reproduction was executed. Last known working revision is unknown.

Evidence and prior checks

Original mount creation custody and retained administrative installation are implemented and published within full Open/openat: lib-ipc 08c3a9d0, ramfsd 84d60c9c, e2fsd 136c1272, fatd 13738d5e, vfsd c2fcb803, posixd 6cbcfe53, docs 962857ba, integration 450c6f2a. The lost mount reply defect is fixed and closed under the testing waiver. Its reproduction remains explicitly source-derived; no executed compiler, VM or fault-injection result is claimed.

Private 0x6FD creation and 0x6FE retirement carry caller-known monotonic keys, complete original correlations, canonical success/refusal shapes and known dispositions. The superseded unowned mount family and consumer adapter are removed. RAM/ext/FAT reserve original IDs and counted mount storage before initialization or media inspection/recovery. Retirement waits for dependent handles, fences stale replay and clears only the original instance. Committed ext journal recovery is not rolled back when private state retires.

Vfsd reserves the namespace record, original route and both correlations before submitting through its counted noncopyable owner. Unknown replies preserve that owner even without a returned mount ID. Administrative installation follows retained raw parent links and earlier actual mount edges, checks actual alias conflicts and covered target type, then publishes in place without losing entering-parent custody. Every failure disables installation before cleanup; refused cleanup retains its original counted row. Independent serving-loop retirement needs neither the initiating client nor fresh correlations. Existing mounts remain available while a failed pending installation drains.

FAT/exFAT supplies original mounted-root/component custody and counted retained ancestry. Retired intermediate caller records become inaccessible to ordinary operations; descendants can retain their entering edges until iterative final release. Dot-dot follows actual held parents and clamps at the original root, including exFAT. Existing identity-replacement exclusion remains effective. Fresh one-use descriptive revisions are consumed before same-turn lookup. POSIX mode/UID/GID/ACL evidence remains unknown; namespace administration does not invent compatibility access policy or acquire additional client authority. Actual compiled arena layouts include the enlarged owner records.

Twenty-one new unit controls are authored and unrun, with three existing actual mount-dispatch controls extended. They cover literal codecs, known refusal and unknown replies, original route/key/correlation cleanup after exhaustion, replay and row reuse, dependent handles, capacity before media work, actual FAT/exFAT ancestry, stale revisions, configured relative/absolute links, aliases, provider crossings, occupied namespace storage and post-creation cleanup refusal. The native realm-binding scenario installs a read-only RAM overlay through an ext symlink parent, opens it through configured and actual namespace spellings, then ascends from a live directory fd. Its explicit budget counts additional namespace and simultaneous client branch pins. No new unit control or VM workload has been executed.

Full Open remains open for conditional creation, authenticated process umask/default ACL inheritance and complete provider policy. No full Open checklist credit is added: 16.91% weighted, 5.21% whole acceptance; 21/460 parents and 441/902 sub-items. Both full EriX guest generations and external Rust/LLVM rebuilding inside EriX remain required and unperformed. Tests, builds, lint checks, documentation generation, VM/profiler execution and CI observation remain deferred. No test pass, warning-free build or measured performance improvement is claimed.

Earlier evidence follows; the original mount custody and retained installation above are current.

Vfsd backend_provider/implementation.rs returns the newly allocated mount ID only after exact reply parsing; namespace.rs mount_fs inserts its vacant record only after that call succeeds. Provider dispatch commits the mount before encoding its result. These source paths establish the missing custody transition. Tests, builds and CI observation are deferred by user instruction; no execution pass or warning-free build is claimed.

Tracking

Track the original-key mount lifetime repair with full Open/openat and the Vfsd implementation PR. Add lost-reply, malformed-reply, allocation exhaustion, cleanup refusal and replay controls, and retain signed coordinated provider/consumer revisions.

## Summary A lost or malformed filesystem-provider mount reply can leave an allocated provider mount without a Vfsd-owned cleanup identity. This is a source-derived reliability and resource-lifetime defect; no executed compiler, fault-injection or VM reproduction is claimed. ## Reproduction Starting from Vfsd revision e08309b1552141a2387390936967ccfb149c9aa7, have a provider commit its mount allocation and suppress or corrupt the corresponding private mount response. Vfsd's mount_provider returns an error before mount_fs inserts the reserved namespace record. The provider mount ID exists only in the lost response. The old mount family supplies neither a caller-known creation key nor an original-key retirement operation. Repeat the fault to consume provider mount capacity. This is a proposed transport fault-injection sequence derived from source, not a claim that it was run. ## Expected and actual behavior Expected: mount creation is counted before submission and retains a caller-known identity that can retire the original provider allocation even if no reply is received. Parent traversal custody must also remain retained until cleanup is positively complete. Actual source behavior: the vacant namespace reservation is abandoned on the fallible mount call; no record retains a possible provider allocation and no private cleanup family can address it without the returned mount ID. ## Environment and identity - Component: Vfsd e08309b1552141a2387390936967ccfb149c9aa7, original private provider mount family 0x620. - Integration source graph: cbca80fc5fb7eb00c783d7e4ff3af1cd9f471890. - Affected provider allocation paths: Ramfsd create_instance and ext/FAT state mount installation after media validation. - Native authority: the existing explicitly delegated filesystem-provider endpoint; no arbitrary client endpoint or supplied numeric ID grants that authority. - No image, VM or hardware reproduction was executed. Last known working revision is unknown. ## Evidence and prior checks Original mount creation custody and retained administrative installation are implemented and published within [full Open/openat](https://git.erikinkinen.fi/erix/posixd/issues/17): [lib-ipc `08c3a9d0`](https://git.erikinkinen.fi/erix/lib-ipc/commit/08c3a9d05ef10b8f1eca48a03bfbb5f4a8abd520), [ramfsd `84d60c9c`](https://git.erikinkinen.fi/erix/ramfsd/commit/84d60c9cac073e6585d057fcd8da89e3d7c21ac0), [e2fsd `136c1272`](https://git.erikinkinen.fi/erix/e2fsd/commit/136c1272d78ea8e0d19bd9e6fa5cd29738ac3927), [fatd `13738d5e`](https://git.erikinkinen.fi/erix/fatd/commit/13738d5e851fe861d8f1ea0aec1f67142a89cf01), [vfsd `c2fcb803`](https://git.erikinkinen.fi/erix/vfsd/commit/c2fcb8036d1656b73fa1e1fdfed55934f618896a), [posixd `6cbcfe53`](https://git.erikinkinen.fi/erix/posixd/commit/6cbcfe537c2bc093823c95a9713cff43e6ac7ad4), [docs `962857ba`](https://git.erikinkinen.fi/erix/docs/commit/962857ba3cb4405dc89b5900b2f38796f555c1bc), [integration `450c6f2a`](https://git.erikinkinen.fi/erix/integration/commit/450c6f2a0c4866ed476375239847c7667395d585). The [lost mount reply defect](https://git.erikinkinen.fi/erix/vfsd/issues/13) is fixed and closed under the testing waiver. Its reproduction remains explicitly source-derived; no executed compiler, VM or fault-injection result is claimed. Private 0x6FD creation and 0x6FE retirement carry caller-known monotonic keys, complete original correlations, canonical success/refusal shapes and known dispositions. The superseded unowned mount family and consumer adapter are removed. RAM/ext/FAT reserve original IDs and counted mount storage before initialization or media inspection/recovery. Retirement waits for dependent handles, fences stale replay and clears only the original instance. Committed ext journal recovery is not rolled back when private state retires. Vfsd reserves the namespace record, original route and both correlations before submitting through its counted noncopyable owner. Unknown replies preserve that owner even without a returned mount ID. Administrative installation follows retained raw parent links and earlier actual mount edges, checks actual alias conflicts and covered target type, then publishes in place without losing entering-parent custody. Every failure disables installation before cleanup; refused cleanup retains its original counted row. Independent serving-loop retirement needs neither the initiating client nor fresh correlations. Existing mounts remain available while a failed pending installation drains. FAT/exFAT supplies original mounted-root/component custody and counted retained ancestry. Retired intermediate caller records become inaccessible to ordinary operations; descendants can retain their entering edges until iterative final release. Dot-dot follows actual held parents and clamps at the original root, including exFAT. Existing identity-replacement exclusion remains effective. Fresh one-use descriptive revisions are consumed before same-turn lookup. POSIX mode/UID/GID/ACL evidence remains unknown; namespace administration does not invent compatibility access policy or acquire additional client authority. Actual compiled arena layouts include the enlarged owner records. Twenty-one new unit controls are authored and unrun, with three existing actual mount-dispatch controls extended. They cover literal codecs, known refusal and unknown replies, original route/key/correlation cleanup after exhaustion, replay and row reuse, dependent handles, capacity before media work, actual FAT/exFAT ancestry, stale revisions, configured relative/absolute links, aliases, provider crossings, occupied namespace storage and post-creation cleanup refusal. The native realm-binding scenario installs a read-only RAM overlay through an ext symlink parent, opens it through configured and actual namespace spellings, then ascends from a live directory fd. Its explicit budget counts additional namespace and simultaneous client branch pins. No new unit control or VM workload has been executed. Full Open remains open for conditional creation, authenticated process umask/default ACL inheritance and complete provider policy. No full Open checklist credit is added: 16.91% weighted, 5.21% whole acceptance; 21/460 parents and 441/902 sub-items. Both full EriX guest generations and external Rust/LLVM rebuilding inside EriX remain required and unperformed. Tests, builds, lint checks, documentation generation, VM/profiler execution and CI observation remain deferred. No test pass, warning-free build or measured performance improvement is claimed. Earlier evidence follows; the original mount custody and retained installation above are current. Vfsd backend_provider/implementation.rs returns the newly allocated mount ID only after exact reply parsing; namespace.rs mount_fs inserts its vacant record only after that call succeeds. Provider dispatch commits the mount before encoding its result. These source paths establish the missing custody transition. Tests, builds and CI observation are deferred by user instruction; no execution pass or warning-free build is claimed. ## Tracking Track the original-key mount lifetime repair with [full Open/openat](https://git.erikinkinen.fi/erix/posixd/issues/17) and the [Vfsd implementation PR](https://git.erikinkinen.fi/erix/vfsd/pulls/3). Add lost-reply, malformed-reply, allocation exhaustion, cleanup refusal and replay controls, and retain signed coordinated provider/consumer revisions.
erikinkinen added this to the Alpha 1 milestone 2026-10-30 09:35:17 +01:00
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
erix/vfsd#13
No description provided.